ratings <- readRDS(file = "data/movieratings.Rds")
str(ratings)

## Classes 'tbl_df', 'tbl' and 'data.frame': 3000 obs. of 6
variables:
## $ userId : int 15 15 15 15 15 15 15 15 15 15 ...
## $ movieId : int 1 32 47 50 110 150 260 296 318 356 ...
## $ rating : num 2 4 5 5 3 3 5 5 2 1 ...
## $ timestamp: int 997938310 997938466 1054449816
997938500 1040205792 997939380 997938437 997938771
997938727 1058250631 ...
## $ title : chr "Toy Story" "Twelve Monkeys" "Seven"
"Usual Suspects, The" ...
## $ genres : Factor w/ 902 levels "(no genres listed)",..:
329 887 889 743 271 453 118 609 700 649 ...

# heatmap
heat <- heatmap(as.matrix((ratings_wide[, -1])), scale = "none", 
    margins = c(12, 5))
row_ordering <- heat$rowInd
col_ordering <- heat$colInd
# reshape long to wide
ratings_wide <- ratings %>% dplyr::select(userId, title, rating) %>% 
    pivot_wider(names_from = title, values_from = rating)
# image wants a matrix of ratings
ratings_mat <- as.matrix(ratings_wide[, -1])

image(t(ratings_mat[row_ordering, col_ordering]))

dist(x, method = "euclidean")

##           1         2         3         4         5
## 2  3.674235                                        
## 3  3.000000  1.870829                              
## 4  7.158911  4.031129  4.387482                    
## 5 10.049876  6.442049  7.483315  4.609772          
## 6 11.747340  8.154753  9.000000  5.315073  2.236068

Smallest distance is between pt 2 and pt 3 (d = 1.87)

dist(x, method = "manhattan")

##      1    2    3    4    5
## 2  6.0                    
## 3  5.0  3.0               
## 4 11.5  5.5  6.5          
## 5 17.0 11.0 12.0  5.5     
## 6 20.0 14.0 15.0  8.5  3.0

x2 <- data.frame(var1 = c(10000, 20000, 5000), var2 = c(10, 40, 
    20))
dist(x2)

##          1        2
## 2 10000.04         
## 3  5000.01 15000.01

dist(x2$var1)

##       1     2
## 2 10000      
## 3  5000 15000

What is the “distance to a group of points”?

… chooses the nearest distance

… chooses the furthest distance

… uses a new point to represent the cluster

Single linkage

Complete linkage

Centroid linkage

same idea – need distance between (potentially) two groups of points

1. Each observation starts in its own cluster

2. While number of clusters > 1

• Merge together the two clusters that are closest to one another
• Update the distance matrix using one of the linkage rules (single, complete, centroid)

# single linkage
hcl_single <- dist(ratings_wide[, -1]) %>% hclust(method = "single")
plot(hcl_single)

# complete linkage
hcl_complete <- dist(ratings_wide[, -1]) %>% hclust(method = "complete")
plot(hcl_complete)

# centroid linkage
hcl_centroid <- dist(ratings_wide[, -1]) %>% hclust(method = "centroid")
plot(hcl_centroid)

user_clusters <- cutree(hcl_complete, h = 9)
ratings_wide <- ratings_wide %>% mutate(user_clusters = user_clusters)
table(ratings_wide$user_clusters)

## 
##  1  2  3  4  5  6 
##  7 30 32 28  2  1

Always check cluster sizes

# transpose and leave out Id and cluster memb columns
ratings_wide_t <- t(ratings_wide[, -c(1, 32)]) %>% as.data.frame()
# column names should be userId's
names(ratings_wide_t) <- ratings_wide$userId
# complete linkage again (or try others)
hcl_complete <- dist(ratings_wide_t) %>% hclust(method = "complete")

title_clusters <- cutree(hcl_complete, k = 6)
table(title_clusters)

## title_clusters
## 1 2 3 4 5 6 
## 9 6 7 5 2 1

# movies in cluster 1
str_c(title_clusters_df[title_clusters_df$title_clus == 1, "title"], 
    collapse = "; ")

## [1] "Toy Story; Braveheart; Apollo 13; Lion King, The;
True Lies; Fugitive, The; Jurassic Park; Aladdin; Dances
with Wolves"

# movies in cluster 2
str_c(title_clusters_df[title_clusters_df$title_clus == 2, "title"], 
    collapse = "; ")

## [1] "Twelve Monkeys; Seven; Terminator 2: Judgment Day;
Matrix, The; Fight Club; Lord of the Rings: , The"

Each observation goes to the cluster with the closest centroid

##      x    y dist1 dist2 col1
## 1 -5.0 -1.0 87.25  0.00  red
## 2 -1.0 -2.0 37.25 17.00  red
## 3 -2.0  3.0 38.25 25.00  red
## 4 -0.4  0.4 20.57 23.12 blue
## 5 -1.5  2.0 30.50 21.25  red
## 6  1.0 -1.5 18.00 36.25 blue

Continue until no change, or some convergence criterion met

1. Randomly choose k initial centroids

2. While some convergence criterion not satisfied

• Allocate each observation to the cluster whose centroid it is closest to
• Recalculate cluster centroids
  
  
• A heuristic for minimizing the sum of within-cluster variances
• Different random starts -> Different clusterings
• How to choose k?

• “By eye”
• Plot % variance explained (between-cluster SS / total SS) against k
• Cross validation
  • Partition data into m parts.
  • For each of the m parts,
    • keep that part aside as test data,
    • fit clustering model on the rest of the data,
    • calculate the sum of the squared distances to the centroids for the test set
  • Choose the value of k that gives the best average performance

df <- ratings_wide %>% dplyr::select_at(vars(-starts_with("user")))
kmeansObj <- kmeans(df, centers = 6)

names(kmeansObj)

## [1] "cluster" "centers" "totss" "withinss"
"tot.withinss"
## [6] "betweenss" "size" "iter" "ifault"

kmeansObj$cluster

## [1] 3 5 3 3 1 5 5 1 4 3 6 4 5 5 3 3 2 4 3 2 5 2 4 6 6 3
5 4 5 3 6 1 4 2 6 1 3
## [38] 1 2 6 2 2 5 2 4 6 3 3 3 6 4 6 1 4 4 5 6 6 1 3 6 2 6
5 3 3 6 2 3 2 3 6 3 3
## [75] 2 6 2 2 3 5 4 5 3 4 1 3 5 5 3 3 6 3 5 5 3 6 6 1 3 2

kmeansObj$centers[, 1:3]

##   Toy Story Twelve Monkeys    Seven
## 1  4.444444       4.000000 4.444444
## 2  4.000000       3.200000 3.500000
## 3  3.571429       3.946429 4.267857
## 4  2.916667       3.000000 3.166667
## 5  3.617647       3.470588 3.676471
## 6  3.105263       3.210526 3.368421

# tidy for ggplot
kmm <- kmeansObj$centers %>% as.data.frame() %>% 
  mutate(user_clus_km = paste0("clust",1:6)) %>%
  pivot_longer(-user_clus_km, names_to = "title", values_to = "rating") 
head(kmm)

## # A tibble: 6 x 3
##   user_clus_km title               rating
##   <chr>        <chr>                <dbl>
## 1 clust1       Toy Story             4.44
## 2 clust1       Twelve Monkeys        4   
## 3 clust1       Seven                 4.44
## 4 clust1       Usual Suspects, The   4.61
## 5 clust1       Braveheart            4.5 
## 6 clust1       Apollo 13             4.44

table(user_clusters, kmeansObj$cluster)

##              
## user_clusters  1  2  3  4  5  6
##             1  0  0  1  0  6  0
##             2  9  2 19  0  0  0
##             3  0 11  8  8  4  1
##             4  0  2  0  2  7 17
##             5  0  0  0  2  0  0
##             6  0  0  0  0  0  1

• transforms a set of observations on possibly correlated variables into observations on a set of uncorrelated variables called principal components.
  
  
• transformation is chosen so each principal component has the largest possible variance subject to the constraint that it is orthogonal to any existing components.

• observation \(i\) on original variables \(\mathbf{x}_i = (x_{i1},x_{i2},\dots,x_{in})\)
  
  
• means for each variable over all observations \(\mathbf{m} = (\bar{x}_{1},\bar{x}_{2},\dots,\bar{x}_{n})\)
  
  
• PC1: linearly transformed observation \(\mathbf{v}_1\mathbf{x}_i = (v_{11}x_{i1},v_{12}x_{i2},\dots,v_{1n}x_{in})\)
  
  
• PC2: linearly transformed observation \(\mathbf{v}_2\mathbf{x}_i = (v_{21}x_{i1},v_{22}x_{i2},\dots,v_{2n}x_{in})\)
  
  
• etc

• Choose coefficients for PC \(j\) \(\mathbf{v}_j\) to maximise \[\displaystyle\frac{1}{m-1}\sum_{i=1}^m(\mathbf{v}^T_j\mathbf{x}_i - \mathbf{v}^T_j\mathbf{m})^2\]

subject to the constraint that \(\mathbf{v}_j\) is orthogonal to all the previous \(\mathbf{v}_j\)’s i.e. \(\mathbf{v}^T_j\mathbf{v}_k=0\) for \(k<j\).

x <- mvrnorm(n = 1000, mu = c(0, 0), Sigma = matrix(c(1, 0.6, 
    0.6, 1), ncol = 2))
plot(x[, 1], x[, 2], xlim = c(-5, 5), ylim = c(-5, 5))

x <- mvrnorm(n = 1000, mu = c(0, 0), Sigma = matrix(c(1, 0.6, 
    0.6, 1), ncol = 2))
plot(x[, 1], x[, 2], xlim = c(-5, 5), ylim = c(-5, 5))

z <- prcomp(x)
plot(z$x[, 1], z$x[, 2])

• Decomposes an (m * n) matrix X into the product of
  • an (m * m) matrix U
  • an (m * n) diagonal matrix D
  • an (n * n) matrix V
    
    
• X = UDV^T
  
  
• Columns of U are left-singular vectors
• Columns of V are right-singular vectors
• Diagonal elements of D are singular values

• SVD is not unique
  
  
• We choose the decomposition that gives the singular values in descending order (there is only one decomposition like this)
  
  
• PCs are equal to the right SVs if you scale the data to have mean 0 and sd 1

# don't need to reorder rows and cols, but will make
# visualizing PCA easier
ratings_scaled_ordered <- scale(ratings_mat)[hcl_row_ordering, 
    hcl_col_ordering]
# do the svd
svd1 <- svd(ratings_scaled_ordered)
# rename svd outputs
u <- svd1$u
v <- svd1$v
d <- diag(svd1$d)
# approximate original data with outer product of first
# singular vector
approx1 <- u[, 1] %*% matrix(d[1, 1], nrow = 1) %*% t(v[, 1])

approx2 <- u[, 1:2] %*% d[1:2, 1:2] %*% t(v[, 1:2])
approx5 <- u[, 1:5] %*% d[1:5, 1:5] %*% t(v[, 1:5])
approx30 <- u %*% d %*% t(v)

PCs are equal to the right SVs if you scale the data to have mean 0 and sd 1

ratings_ordered <- ratings_mat[hcl_row_ordering, hcl_col_ordering]
pca1 <- prcomp(ratings_ordered, scale = TRUE)
plot(pca1$rotation[, 1], svd1$v[, 1])

# transform into matrix form
mona_mat <- as.data.frame(mona) %>% pivot_wider(names_from = y, 
    values_from = value) %>% dplyr::select(-x) %>% as.matrix()
image(mona_mat[, nrow(mona_mat):1], col = gray.colors(100))

# svd
svd1 <- svd(mona_mat)
u <- svd1$u
v <- svd1$v
d <- diag(svd1$d)

# number of singular values
nsv <- 1
# approximate original data with outer product of first N
# singular vectors
approx <- u[, 1:nsv] %*% matrix(d[1:nsv, 1:nsv], nrow = nsv) %*% 
    t(v[, 1:nsv])